Amyloid PET Quantification Via End-to-End Training of a Deep Learning

• Published in: Nuclear medicine and molecular imaging Vol. 53; no. 5; pp. 340 - 348
• Main Authors: Kim, Ji-Young, Suh, Hoon Young, Ryoo, Hyun Gee, Oh, Dongkyu, Choi, Hongyoon, Paeng, Jin Chul, Cheon, Gi Jeong, Kang, Keon Wook, Lee, Dong Soo
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.10.2019; Springer Nature B.V; 대한핵의학회
• Subjects: Cardiology; Deep learning; Emission analysis; Evaluation; Image acquisition; Imaging; Machine learning; Medical imaging; Medicine; Medicine & Public Health; Neurology; Nuclear Medicine; Oncology; Original; Original Article; Orthopedics; Positron emission; Radioactive tracers; Radiology; Test sets; Tomography; Training; 방사선과학; Deep learning; Alzheimer’s disease; Quantification; Amyloid PET; Convolutional neural network
• ISSN: 1869-3474; 1869-3482
• DOI: 10.1007/s13139-019-00610-0

Purpose Although quantification of amyloid positron emission tomography (PET) is important for evaluating patients with cognitive impairment, its routine clinical use is hampered by complicated preprocessing steps and required MRI. Here, we suggested a one-step quantification based on deep learning using native-space amyloid PET images of different radiotracers acquired from multiple centers. Methods Amyloid PET data of the Alzheimer Disease Neuroimaging Initiative (ADNI) were used for this study. A training/validation consists of 850 florbetapir PET images. Three hundred sixty-six florbetapir and 89 florbetaben PET images were used as test sets to evaluate the model. Native-space amyloid PET images were used as inputs, and the outputs were standardized uptake value ratios (SUVRs) calculated by the conventional MR-based method. Results The mean absolute errors (MAEs) of the composite SUVR were 0.040, 0.060, and 0.050 of training/validation and test sets for florbetapir PET and a test set for florbetaben PET, respectively. The agreement of amyloid positivity measured by Cohen’s kappa for test sets of florbetapir and florbetaben PET were 0.87 and 0.89, respectively. Conclusion We suggest a one-step quantification method for amyloid PET via a deep learning model. The model is highly reliable to quantify the amyloid PET regardless of multicenter images and various radiotracers.